Minor identification and risk warning system in special scene based on probability confidence

Abstract

This invention discloses a system for identifying and warning of minors in special scenarios based on probability confidence, relating to the field of data analysis. It includes: a data acquisition module, used to dynamically adjust parameters according to the ambient light at the entrance of the system-configured location and the shooting angle, to acquire facial and body feature image data of the object to be identified; and an extraction module, used to receive the image data acquired by the acquisition module, extract facial proportions, facial contours, body lines, facial skin texture, and body proportion information from the image data, and construct a feature vector set. This invention dynamically optimizes acquisition parameters based on ambient light and shooting angle to ensure image data quality, comprehensively extracts multi-dimensional key information such as facial proportions, facial contours, and body lines, and quantifies and constructs feature vectors. Through dynamic weight allocation, it strengthens the role of high-discrimination features, and combined with a probability model of minor characteristics, it achieves accurate matching calculations and outputs a reliable probability confidence value.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a system for identifying and warning of minors in special scenarios based on probability confidence. Background Technology

[0002] Currently, most adult entertainment venues rely on security personnel to control the entry of consumers. However, law enforcement officers can only make a simple assessment of whether consumers are adults through facial observation, making it extremely difficult to prevent minors from entering such entertainment venues that are not suitable for minors.

[0003] The utility model patent application with application number 202322335159.6 discloses a minor identification and protection early warning system for entertainment venues. The system aims to solve the problems of "verification through identity verification equipment requires the person to show their ID card for verification. On the one hand, the person may not carry their ID card at all times, and on the other hand, the supervision of identity verification is loose and easy to forge; verification through facial recognition gate equipment is large, inconvenient to install, and has low security. Personnel may forcibly pass through to enter the entertainment venue."

[0004] However, traditional facial recognition is mainly used for identity verification and requires the support of a backend database. For minors, the database mainly contains images of their faces from childhood. Due to significant changes in appearance during growth and large individual differences, the recognition rate is even lower. In addition, interference from external factors can affect the accurate identification of facial attributes.

[0005] To address this, we propose a system for identifying minors and providing risk warnings in special scenarios based on probabilistic confidence. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a minor identification and risk warning system based on probability confidence in special scenarios, which can effectively solve the problems of the existing technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions;

[0008] This invention discloses a system for identifying minors and providing risk warnings in special scenarios based on probability confidence, comprising:

[0009] The system comprises the following modules: Acquisition module, Extraction module, and Calculation module. Acquisition module dynamically adjusts parameters based on ambient light and shooting angle at the entrance of the venue to acquire facial and body feature image data of the target object. Extraction module receives the image data acquired by the acquisition module and extracts facial proportions, facial contours, body lines, facial skin texture, and body proportion information to construct a feature vector set. Calculation module assigns dynamic weights to the feature vector set and performs adaptive matching calculations with a preset underage feature probability model to output a confidence value for the underage identity probability. Judgment module dynamically adjusts the judgment threshold based on the venue's time of day and crowd density, compares the probability confidence value with the dynamic threshold, and generates a result indicating a suspected underage identity. Warning module classifies warning levels based on probability confidence values ​​and outputs a risk warning signal carrying the confidence value and warning level to the venue management terminal. Feedback module collects the corresponding feature vector sets of those verified as underage after a warning, stores them in a structured manner according to feature dimensions to construct a sample library, and feeds the sample library back into the underage feature probability model to iterate the model's parameters.

[0010] The acquisition module is interconnected with the extraction module via a wireless network. The extraction module is interconnected with the computing module via a wireless network. The computing module is interconnected with the judgment module via a wireless network. The judgment module is interconnected with the early warning module via a wireless network. The early warning module is interconnected with the feedback module via a wireless network.

[0011] Furthermore, the dynamic adjustment parameters of the acquisition module include:

[0012] Acquire real-time brightness and color temperature values ​​of the ambient light at the entrance of the venue, as well as the real-time pitch and horizontal tilt angles of the shooting device, to adjust the exposure time, ISO, and focal length:

[0013] ;

[0014] In the formula: This is the adjusted exposure time; The preset baseline exposure time; , , For preset weighting coefficients, , , The sum is 1; Preset reference brightness; This is the real-time brightness value; For real-time pitch angle; Preset reference pitch angle; This is the real-time color temperature value; Preset reference color temperature; , , Preset adjustment index; This is the adjusted ISO value; Preset baseline ISO value; Adjust the index for the preset ISO; To adjust the focal length; The preset reference focal length; This is the real-time horizontal deflection angle; This is the preset baseline horizontal deflection angle.

[0015] Furthermore, in the feature vector set construction stage of the extraction module, the extracted facial proportions, facial contours, body lines, facial skin texture, and body proportion information are quantified respectively:

[0016] Facial proportion quantification: facial width-to-height ratio, facial feature spacing ratio, which is the ratio of the distance between the eyes to the sum of the width of the nose and lips;

[0017] Facial contour quantification: eyebrow curvature, taking the average curvature of the eyebrow curve; eye shape eccentricity, taking the ratio of the major semi-axis to the minor semi-axis of the eye contour ellipse.

[0018] Quantification of body shape: shoulder-to-waist ratio, which is the ratio of shoulder width to waist circumference; leg-to-body ratio, which is the ratio of leg length to height.

[0019] Facial skin texture quantization: texture orientation entropy, which is obtained by taking the orientation distribution entropy calculated based on the gray-level co-occurrence matrix; texture contrast, which is obtained by taking the contrast feature value of the gray-level co-occurrence matrix.

[0020] Body proportion quantification: head-to-body ratio;

[0021] Based on the quantized feature values ​​described above, construct the feature vector:

[0022] ,in Let be the initial weights for each feature, and .

[0023] Furthermore, the dynamic weight allocation of the calculation module follows the following rules:

[0024] ;

[0025] In the formula: The dynamic value of the i-th weight; Let be the inter-class variance of the i-th feature between the minor and non-minor samples. Similarly; Let be the within-class variance of the i-th feature in the sample of minors. Similarly; The preset time decay factor; This represents the time interval between the current moment and the last time the discriminative power of this feature was updated;

[0026] After assigning dynamic weights, the feature vector V is compared with the dynamic weight vector. Perform a dot product to obtain a weighted eigenvector. Then The input is a probability model of minor features, which is used for adaptive matching. The output is the probability confidence value P.

[0027] Furthermore, when the extraction module extracts body contour information, it performs edge detection on the collected body feature image data to obtain a continuous curve set of body contours, and then calculates the rate of curvature change of each curve segment in the curve set. Where K is the curvature of the curve segment and s is the arc length of the curve segment, in order to filter out the rate of change of curvature. Curve segments within a preset range are used as key body posture lines;

[0028] At the same time, the length ratio of key body lines is calculated, that is, the ratio of the total length of key body lines to the total perimeter of the body contour, and then this ratio is added to the corresponding body line component in the feature vector V.

[0029] Furthermore, the dynamic judgment threshold adjustment process of the judgment module is as follows:

[0030] Obtain the current time period type M and real-time crowd density D of the current location, where M=1 corresponds to peak time, M=2 corresponds to off-peak time, and M=3 corresponds to low-peak time.

[0031] Calculate the dynamic judgment threshold ;

[0032] In the formula: The preset benchmark threshold is used for judgment. , The preset coefficient corresponding to time period type M; Real-time pedestrian density; Preset baseline pedestrian density;

[0033] The expression for the probabilistic model of minor characteristics is: ;

[0034] In the formula: The weighted cosine similarity between the feature vector and the reference vector is denoted as . , Represents the weighted eigenvector, This represents the baseline vector of the pre-trained minor feature model of the system. The L2 norm of a vector. Indicates the preset minimum value; As a dynamic weighted effectiveness factor, , represent the standard deviation of the dynamic weight vector Ω and the maximum value in the dynamic weight vector, respectively; The sample distribution adaptation factor is the ratio of the number of matches between the current feature vector and the same type of sample in the model to the total number of the same type of sample in the model. This is the distance attenuation coefficient;

[0035] The probability confidence value P and the dynamic threshold Comparison: If P ≥ If P < 0, then it is determined to be a suspected minor; if P < 0. If so, it is determined that the individual is not a suspected minor.

[0036] Furthermore, the warning level classification logic of the warning module is as follows:

[0037] Obtain feature matching stability ,in Let N be the standard deviation of the probability confidence values ​​obtained from N consecutive data collections. The mean of the N confidence scores is the average of the N scores, where N represents the preset number of data collections.

[0038] Calculate comprehensive early warning indicators ,in The preset adjustment coefficient, This represents the probability confidence level.

[0039] when When the value is not less than the preset high-level threshold, it is a Level 1 warning; when... When the threshold is not less than the preset medium-level threshold and less than the preset high-level threshold, it is a level two warning; when When the value is less than the preset threshold, it is a Level 3 warning; the warning module will output a risk warning signal carrying P, S, Q and the warning level to the venue management terminal.

[0040] Furthermore, the sample library construction of the feedback module conforms to:

[0041] The feature vector set verified as belonging to minors is divided into a facial feature subset and a body posture feature subset according to feature dimensions;

[0042] For each subset, calculate the contribution of the feature dimension. ,in This represents the mean of that dimension in the dynamic weights. This represents the average contribution probability of this dimension in the matching operation;

[0043] Based on contribution Build a hierarchical sample library for the index: use the feature dimensions with a contribution value higher than the preset contribution value threshold as the first-level index, and the rest as the second-level index.

[0044] Furthermore, the model feedback of the feedback module follows the following:

[0045] Based on the index hierarchy of the sample library, feature samples corresponding to the first-level index are selected first for model parameter iteration. The iteration update formula is as follows: ;

[0046] In the formula: These are the updated model parameters; These are the original parameters of the model; Set the learning rate; The gradient of the model's loss function; This represents the average contribution of the corresponding index-level feature dimension.

[0047] in, The preset value range is [0.001, 0.1]. The value is larger when the parameters deviate greatly from the optimal solution in the early stage of model feedback and the average feature contribution is high. The value is smaller when the model approaches the optimal solution in the later stage, the average feature contribution is low, or the gradient of the loss function is small.

[0048] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects:

[0049] This invention provides a system for identifying and warning of minors in special scenarios based on probability confidence. During operation, the system dynamically optimizes acquisition parameters based on ambient light and shooting angle to ensure image data quality. It comprehensively extracts key information from multiple dimensions, including facial proportions, facial contours, and body lines, and quantifies and constructs feature vectors. Dynamic weight allocation strengthens the role of highly discriminative features, and combined with a probability model of minor characteristics, it achieves accurate matching calculations and outputs reliable probability confidence values. Simultaneously, it flexibly adjusts the judgment threshold based on location, time of day, and crowd density to reduce recognition errors caused by scene differences. Furthermore, it classifies warning levels based on probability confidence and matching stability, making risk alerts more targeted. It also collects verified valid samples, stores them in a structured manner according to feature dimensions, and uses this data to optimize model parameters, continuously improving recognition adaptability and accuracy, effectively reducing the probability of misjudgments and missed judgments, and providing efficient and intelligent governance for risk prevention and control of minors in special scenarios. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0051] Figure 1This is a schematic diagram of the structure of a minor identification and risk warning system in a special scenario based on probability confidence. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0053] The present invention will be further described below with reference to embodiments.

[0054] Example:

[0055] This embodiment presents a probabilistic confidence-based system for minor identification and risk warning in special scenarios, such as... Figure 1 As shown, it includes:

[0056] The acquisition module is used to dynamically adjust parameters based on the ambient light at the entrance of the venue and the shooting angle configured by the system to acquire facial and body feature image data of the object to be identified;

[0057] The dynamically adjustable parameters of the acquisition module include:

[0058] Acquire real-time brightness and color temperature values ​​of the ambient light at the entrance of the venue, as well as the real-time pitch and horizontal tilt angles of the shooting device, to adjust the exposure time, ISO, and focal length:

[0059] ;

[0060] In the formula: This is the adjusted exposure time; The preset baseline exposure time; , , For preset weighting coefficients, , , The sum is 1; Preset reference brightness; This is the real-time brightness value; For real-time pitch angle; Preset reference pitch angle; This is the real-time color temperature value; Preset reference color temperature; , , Preset adjustment index; This is the adjusted ISO value; Preset baseline ISO value; Adjust the index for the preset ISO; To adjust the focal length; The preset reference focal length; This is the real-time horizontal deflection angle; The preset reference horizontal deflection angle;

[0061] The above formula combines the real-time brightness and color temperature values ​​of the ambient light at the entrance of the venue with the real-time pitch and horizontal tilt angles of the shooting device. By setting weight coefficients and adjustment indices, it constructs a linkage adjustment logic for exposure time, ISO and focal length. The weight coefficients are dynamically allocated according to the degree of influence of the environment and equipment status on image quality, and the adjustment index is adapted to the parameter response sensitivity under different impact scenarios. This achieves accurate matching between the collected parameters and the real-time scene, ensuring the clarity and effectiveness of feature image data in complex environments.

[0062] in, The preset value range is (0, 1). When the ambient brightness fluctuates greatly and has a significant impact on image brightness and noise control, the value is larger. When the ambient brightness is stable and has little impact on the captured image quality, the value is smaller. The value range is preset to (0, 1). When the pitch angle of the shooting device changes frequently and the offset is large, affecting the clarity of feature capture, the value is larger. When the pitch angle is stable and the offset is small, the value is smaller. The value range is preset to (0, 1). When the color temperature of the light source fluctuates drastically and has a significant impact on the color reproduction of the image, the value is larger. When the color temperature is stable and the risk of color distortion is low, the value is smaller. The preset value range is [0.5, 3.0]. The larger the value is when the change in ambient brightness has a greater impact on the image brightness compliance rate and when a fast response adaptation of the exposure time is required, the smaller the value is when the change in brightness has a smaller impact on the image quality. The preset value range is [0.3, 2.5]. The larger the value is when the shooting pitch angle offset has a more significant impact on the depth of field and feature sharpness of the image, and when targeted compensation of exposure time is required, the smaller the value is when the offset has a smaller impact on image quality. The preset value range is [0.4, 2.8]. The larger the value, the greater the impact of color temperature change on the accuracy of image color reproduction and the more necessary it is to optimize color performance with exposure time. The smaller the value, the smaller the impact of color temperature change on color recognition. The preset value range is [0.6, 3.2]. The larger the value is when the change in exposure time causes a greater deviation in image brightness and the ISO parameter needs to be quickly linked to correct the brightness, the smaller the value is when the change in exposure time has a smaller impact on brightness.

[0063] The extraction module is used to receive image data collected by the acquisition module, extract facial proportions, facial features, body lines, facial skin texture, and body proportion information from the image data, and construct a feature vector set.

[0064] In the feature vector set construction stage of the extraction module, the extracted facial proportions, facial contours, body lines, facial skin texture, and body proportion information are quantified respectively:

[0065] Facial proportion quantification: facial width-to-height ratio, facial feature spacing ratio, which is the ratio of the distance between the eyes to the sum of the width of the nose and lips;

[0066] Facial contour quantification: eyebrow curvature, taking the average curvature of the eyebrow curve; eye shape eccentricity, taking the ratio of the major semi-axis to the minor semi-axis of the eye contour ellipse.

[0067] Quantification of body shape: shoulder-to-waist ratio, which is the ratio of shoulder width to waist circumference; leg-to-body ratio, which is the ratio of leg length to height.

[0068] Facial skin texture quantization: texture orientation entropy, which is obtained by taking the orientation distribution entropy calculated based on the gray-level co-occurrence matrix; texture contrast, which is obtained by taking the contrast feature value of the gray-level co-occurrence matrix.

[0069] Body proportion quantification: head-to-body ratio;

[0070] Based on the quantized feature values ​​described above, construct the feature vector:

[0071] ,in Let be the initial weights for each feature, and ;

[0072] The calculation module is used to assign dynamic weights to the feature vector set, perform adaptive matching calculations with the preset minor feature probability model, and output the confidence value of the minor's identity probability.

[0073] The dynamic weight allocation of the calculation module follows the following rules:

[0074] ;

[0075] In the formula: The dynamic value of the i-th weight; Let be the inter-class variance of the i-th feature between the minor and non-minor samples. Similarly; Let be the within-class variance of the i-th feature in the sample of minors. Similarly; The preset time decay factor; This represents the time interval between the current moment and the last time the discriminative power of this feature was updated;

[0076] The above formula uses the ratio of inter-class variance to intra-class variance of features to reflect feature discrimination. It introduces a time decay factor to consider the timeliness of feature discrimination. Through normalization, the dynamic weight of each feature is obtained, so that features with high discrimination and strong timeliness are given higher weights. It automatically adapts to the dynamic changes in feature importance, thereby improving the accuracy of feature vectors in representing the identity of minors.

[0077] After assigning dynamic weights, the feature vector V is compared with the dynamic weight vector. Perform a dot product to obtain a weighted eigenvector. Then The input is a probability model of minor characteristics, which is used for adaptive matching. The output is the probability confidence value P.

[0078] The samples from minors and non-minors are the initial labeled sample sets preset by the system, and are dynamically updated through the feedback module. The preset value range is [0.01, 0.1]. The larger the value, the more significant the change in feature discrimination over time; the smaller the value, the more gradual the change in feature discrimination over time.

[0079] When extracting body contour information, the extraction module performs edge detection on the collected body feature image data to obtain a continuous curve set of body contours, and then calculates the rate of curvature change of each curve segment in the curve set. Where K is the curvature of the curve segment and s is the arc length of the curve segment, in order to filter out the rate of change of curvature. Curve segments within a preset range are used as key body posture lines;

[0080] At the same time, the length ratio of key body lines is calculated, that is, the ratio of the total length of key body lines to the total perimeter of the body contour, and then this ratio is added to the corresponding body line component in the feature vector V;

[0081] The judgment module is used to dynamically adjust the judgment threshold based on the location, time period, and crowd density, compare the probability confidence value with the dynamic threshold, and generate a judgment result for suspected minors.

[0082] The dynamic threshold adjustment process of the judgment module is as follows:

[0083] Obtain the current time period type M and real-time crowd density D of the current location, where M=1 corresponds to peak time, M=2 corresponds to off-peak time, and M=3 corresponds to low-peak time.

[0084] Calculate the dynamic judgment threshold ;

[0085] In the formula: The preset benchmark threshold is used for judgment. , The preset coefficient corresponding to time period type M; Real-time pedestrian density; Preset baseline pedestrian density;

[0086] The above formula is based on the type of time period and real-time crowd density. By linking the preset coefficient corresponding to the time period with the proportion of crowd density, the benchmark judgment threshold is dynamically adjusted. During periods of high crowd density, the threshold is increased to adapt to high recognition pressure. During periods of high risk sensitivity, the coefficient is optimized to enhance recognition sensitivity, ensuring that the judgment standard is dynamically adapted to the actual scene of the venue, so as to balance recognition accuracy and efficiency.

[0087] The expression for the probabilistic model of minor characteristics is: ;

[0088] In the formula: The weighted cosine similarity between the feature vector and the reference vector is denoted as . , Represents the weighted eigenvector, This represents the baseline vector of the pre-trained minor feature model of the system. The L2 norm of a vector. Indicates the preset minimum value; As a dynamic weighted effectiveness factor, , represent the standard deviation of the dynamic weight vector Ω and the maximum value in the dynamic weight vector, respectively; The sample distribution adaptation factor is the ratio of the number of matches between the current feature vector and the same type of sample in the model to the total number of the same type of sample in the model. This is the distance attenuation coefficient;

[0089] The above formula integrates the modified cosine similarity between the weighted feature vector and the baseline vector, the dynamic weight effectiveness factor, and the sample distribution adaptation factor. At the same time, it introduces the distance decay coefficient to weaken the influence of samples with excessively large vector differences. It calculates the probability confidence value by comprehensively considering multiple dimensions of indicators, taking into account both the degree of feature matching and the effectiveness of weights and the adaptability of sample distribution, and ultimately improving the reliability of minor identity determination.

[0090] The probability confidence value P and the dynamic threshold Comparison: If P ≥ If P < 0, then it is determined to be a suspected minor; if P < 0. If so, the individual is determined not to be a suspected minor;

[0091] in, , The value range is (0,1), and the higher the population density during a given time period, the better. The larger the value, the lower the population density during that time period. The smaller the value, the higher the sensitivity requirement for risk identification over a given period. The larger the value, the lower the sensitivity requirement for risk identification over a given period. The smaller the value, The preset value range is (0.8, 3.2). When the sample distribution dispersion between the feature vector and the reference vector is higher... The larger the value, the more concentrated the sample distribution. The smaller the value;

[0092] The early warning module is used to classify early warning levels based on probability confidence values ​​and output risk warning signals carrying confidence values ​​and early warning levels to the site management terminal.

[0093] The logic for classifying early warning levels in the early warning module is as follows:

[0094] Obtain feature matching stability ,in Let N be the standard deviation of the probability confidence values ​​obtained from N consecutive data collections. The mean of the N confidence scores is the average of the N scores, where N represents the preset number of data collections.

[0095] Calculate comprehensive early warning indicators ,in The preset adjustment coefficient, This represents the probability confidence level.

[0096] The above formula calculates the feature matching stability by using the standard deviation and mean of N consecutive confidence values. It combines the probability confidence value with the preset adjustment coefficient to construct a comprehensive early warning index. The adjustment coefficient is dynamically adapted according to the difference between the confidence value and the stability. The early warning level is divided according to the index range, which reflects both the credibility of identity determination and the stability of the matching result.

[0097] when When the value is not less than the preset high-level threshold, it is a Level 1 warning; when... When the threshold is not less than the preset medium-level threshold and less than the preset high-level threshold, it is a level two warning; when When the value is less than the preset threshold, it is a Level 3 warning; the warning module will output a risk warning signal carrying P, S, Q and the warning level to the venue management terminal;

[0098] in, The value range is preset to (0,1]. The larger the difference between the probability confidence value and the feature matching stability, the larger the value; the closer P and S are, the smaller the value.

[0099] The feedback module is used to collect the corresponding feature vector sets of minors after the warning is verified, classify and store them in a structured manner according to feature dimensions to build a sample library, and feed the sample library back into the minor feature probability model to iterate the parameters of the minor feature probability model.

[0100] The sample library construction of the feedback module follows the following rules:

[0101] The feature vector set verified as belonging to minors is divided into a facial feature subset and a body posture feature subset according to feature dimensions;

[0102] For each subset, calculate the contribution of the feature dimension. ,in This represents the mean of that dimension in the dynamic weights. This represents the average contribution probability of this dimension in the matching operation, which is characterized by the statistical mean of the contribution of this feature dimension to the probability confidence value after dynamic weighting in historical validation samples.

[0103] Based on contribution To build a hierarchical sample library for the index: feature dimensions with a contribution higher than a preset contribution threshold are used as the first-level index, and the rest are used as the second-level index;

[0104] The feedback module's model feedback follows:

[0105] Based on the index hierarchy of the sample library, feature samples corresponding to the first-level index are selected first for model parameter iteration. The iteration update formula is as follows: ;

[0106] In the formula: These are the updated model parameters; These are the original parameters of the model; Set the learning rate; The gradient of the model's loss function; This represents the average contribution of the corresponding index-level feature dimension.

[0107] The above formula takes high-contribution feature samples from the first-level index of the sample library as input, and introduces the learning rate and the mean of feature contribution to adjust the parameter update magnitude. The learning rate adapts to the parameter optimization needs at different stages of model iteration, and the mean of contribution highlights the role of high-value features in model updates. Combined with the cross-entropy loss function and its gradient, the model parameters are accurately iterated, continuously improving the model's ability to fit the features of minors.

[0108] in, The preset value range is [0.001, 0.1]. The value is larger when the parameters deviate greatly from the optimal solution in the early stage of model feedback and the average feature contribution is high. The value is smaller when the model approaches the optimal solution in the later stage, the average feature contribution is low, or the gradient of the loss function is small.

[0109] The weighted feature vector of the feature samples corresponding to the first-level index The model takes the true labels of the samples as input and the true labels of the samples as the target output, i.e., minors are labeled as 1 and non-minors are labeled as 0, and a loss function is constructed accordingly. , This indicates the number of first-level index feature samples participating in the iteration. This represents the true label of the k-th sample. This represents the model's predicted probability for the k-th sample. This represents the mean contribution of the feature dimension corresponding to the k-th sample;

[0110] The above formula is based on the true labels and predicted probabilities of the first-level index feature samples. It introduces the weighted average contribution of the feature dimension corresponding to the sample to calculate the loss value. This not only follows the core logic of the loss function of the classification task, but also strengthens the role of high-value samples in model training by means of the average contribution, guides the model to focus on learning the key feature patterns for identification, and improves the efficiency and effectiveness of model iteration.

[0111] but , where T is the vector transpose operator, m represents the total number of parameters to be updated in the model, and the calculation results of each partial derivative together constitute the gradient vector of the model loss function;

[0112] The data acquisition module interacts with the extraction module via a wireless network. The extraction module interacts with the computing module via a wireless network. The computing module interacts with the judgment module via a wireless network. The judgment module interacts with the early warning module via a wireless network. The early warning module interacts with the feedback module via a wireless network.

[0113] In this embodiment, the acquisition module dynamically adjusts parameters based on the ambient light and shooting angle at the entrance of the venue configured by the system to acquire facial and body feature image data of the object to be identified. The extraction module then receives the image data acquired by the acquisition module and extracts facial proportions, facial contours, body lines, facial skin texture, and body proportion information from the image data to construct a feature vector set. The calculation module then assigns dynamic weights to the feature vector set and performs adaptive matching calculations with a preset minor feature probability model to output a probability confidence value for minor identity. The judgment module further dynamically adjusts the judgment threshold based on the venue time and crowd density, compares the probability confidence value with the dynamic threshold, generates a suspected minor judgment result, and the early warning module classifies the early warning level according to the probability confidence value and outputs a risk warning signal carrying the confidence value and warning level to the venue management terminal. Finally, the feedback module collects the corresponding feature vector sets of minors verified after the early warning, classifies and stores them in a structured manner according to feature dimensions to construct a sample library, and feeds the sample library back into the minor feature probability model to iterate the parameters of the minor feature probability model.

[0114] In the above embodiments, the system can dynamically adapt to the environment and shooting conditions, accurately capture relevant features and intelligently match and analyze them. It can flexibly adjust the judgment criteria based on time period and crowd flow, output warnings in a graded manner, and continuously optimize the recognition accuracy through actual cases. It can efficiently identify suspected minors, help venues accurately prevent and control risks, protect the safety of minors, and improve management efficiency and standardization.

[0115] Referring to the system in the above embodiments, an application example of the system is shown:

[0116] As a special venue, XX Internet Cafe requires strict control over minors' entry and has deployed this identification and risk warning system. When the system is running, the data acquisition module first obtains real-time ambient light brightness and color temperature values ​​at the entrance of the internet cafe, as well as the pitch and horizontal angles of the shooting equipment. Based on this data, the system automatically adjusts the parameters, ultimately determining an exposure time of 0.04 seconds, an ISO value of 320, and a focal length of 55mm, accurately capturing facial and body feature images of those entering the cafe.

[0117] After receiving the acquired image data, the extraction module quantifies the relevant features: the facial aspect ratio is 1.3; the ratio of inter-eye distance to the sum of nose and lip width is 0.4; the mean curvature of the eyebrow curve is 0.6; the ratio of the major to minor axis of the eye contour ellipse is 1.8; the ratio of shoulder width to waist circumference is 1.5; the ratio of leg length to height is 0.6; the texture orientation entropy calculated based on the gray-level co-occurrence matrix is ​​1.2; the contrast feature value is 0.8; and the head-to-body ratio is 6.5. Based on these quantization results, the system constructs the corresponding feature vector.

[0118] The calculation module assigns dynamic weights based on factors such as the inter-class variance, intra-class variance, and time interval of each feature. It then performs a dot product between the feature vector and the dynamic weight vector to obtain a weighted feature vector, which is then input into the minor feature probability model for matching operations. Finally, it outputs a confidence value of 0.85 for the probability of the person being a minor.

[0119] The judgment module identifies that the current period is peak operating time for the internet cafe (M=1), and the real-time crowd density is higher than the preset baseline crowd density. Combining these two factors, the dynamic judgment threshold is adjusted, and the calculated dynamic threshold is 0.78. Since the probability confidence value of 0.85 is greater than the dynamic threshold of 0.78, the system determines that the person is a suspected minor.

[0120] The early warning module then calculates the feature matching stability. The standard deviation of the confidence values ​​collected for five consecutive times is 0.03, and the mean is 0.84, resulting in a feature matching stability of 0.96. Combined with the preset adjustment coefficient of 0.8, the comprehensive early warning index Q is calculated to be 0.92. According to the system settings, when Q reaches the high-level threshold, a level one early warning is triggered. A risk warning signal carrying a confidence value of 0.85, a feature matching stability of 0.96, a comprehensive early warning index of 0.92, and a level one early warning level is sent to the internet cafe management terminal. The management personnel promptly verify the identity of the person.

[0121] Upon verification, the individual was indeed a minor. The feedback module divided the corresponding feature vector set into subsets based on facial and body features, calculated the contribution of each feature dimension, and used facial and body proportion features with contribution values ​​exceeding a preset threshold as primary indices, while the rest were used as secondary indices to construct a hierarchical sample library. Subsequently, the system prioritized selecting feature samples corresponding to the primary indices, and at a preset learning rate of 0.01, iteratively updated the parameters of the minor's feature probability model by combining the gradient of the model's loss function and the average contribution of the feature dimensions.

[0122] In summary, the system in the above embodiments dynamically optimizes the acquisition parameters based on ambient light and shooting angle to ensure image data quality. It comprehensively extracts key information from multiple dimensions such as facial proportions, facial contours, and body lines, and quantifies and constructs feature vectors. Through dynamic weight allocation, it strengthens the role of high-discrimination features. Combined with a probabilistic model of minors' characteristics, it achieves accurate matching calculations and outputs reliable probability confidence values. At the same time, it flexibly adjusts the judgment threshold according to the location, time of day, and crowd density to reduce recognition errors caused by scene differences. Furthermore, it divides the warning level by probability confidence and matching stability, making risk warnings more targeted. It can also collect and validated valid samples, store them in a structured manner according to feature dimensions, and feed back to optimize model parameters, continuously improving recognition adaptability and accuracy, effectively reducing the probability of misjudgment and missed judgment, and providing efficient and intelligent governance for risk prevention and control of minors in special scenarios.

[0123] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A system for identifying and warning of minors in special scenarios based on probability confidence, characterized in that, include: The acquisition module is used to dynamically adjust parameters based on the ambient light at the entrance of the venue and the shooting angle configured by the system to acquire facial and body feature image data of the object to be identified; The extraction module is used to receive image data collected by the acquisition module, extract facial proportions, facial features, body lines, facial skin texture, and body proportion information from the image data, and construct a feature vector set. The calculation module is used to assign dynamic weights to the feature vector set, perform adaptive matching calculations with the preset minor feature probability model, and output the confidence value of the minor's identity probability. The judgment module is used to dynamically adjust the judgment threshold based on the location, time period, and crowd density, compare the probability confidence value with the dynamic threshold, and generate a judgment result for suspected minors. The early warning module is used to classify early warning levels based on probability confidence values ​​and output risk warning signals carrying confidence values ​​and early warning levels to the site management terminal. The feedback module is used to collect the corresponding feature vector sets of minors after the early warning is verified. The vectors are classified and stored in a structured manner according to the feature dimension to build a sample library. The sample library is fed back into the minor feature probability model to iterate the parameters of the minor feature probability model.

2. The system for identifying and warning of minors in special scenarios based on probability confidence as described in claim 1, characterized in that, The dynamic adjustment parameters of the acquisition module include: Acquire real-time brightness and color temperature values ​​of the ambient light at the entrance of the venue, as well as the real-time pitch and horizontal tilt angles of the shooting device, to adjust the exposure time, ISO, and focal length: ； In the formula: This is the adjusted exposure time; The preset baseline exposure time; , , For preset weighting coefficients, , , The sum is 1; Preset reference brightness; This is the real-time brightness value; For real-time pitch angle; Preset reference pitch angle; This is the real-time color temperature value; Preset reference color temperature; , , Preset adjustment index; This is the adjusted ISO value; Preset baseline ISO value; Adjust the index for the preset ISO; To adjust the focal length; The preset reference focal length; This is the real-time horizontal deflection angle; This is the preset baseline horizontal deflection angle.

3. The system for identifying and warning of minors in special scenarios based on probability confidence as described in claim 1, characterized in that, In the feature vector set construction stage of the extraction module, the extracted facial proportions, facial contours, body lines, facial skin texture, and body proportions are quantified respectively: Facial proportion quantification: facial width-to-height ratio, facial feature spacing ratio, which is the ratio of the distance between the eyes to the sum of the width of the nose and lips; Facial contour quantification: eyebrow curvature, taking the average curvature of the eyebrow curve; eye shape eccentricity, taking the ratio of the major semi-axis to the minor semi-axis of the eye contour ellipse. Quantification of body shape: shoulder-to-waist ratio, which is the ratio of shoulder width to waist circumference; leg-to-body ratio, which is the ratio of leg length to height. Facial skin texture quantization: texture orientation entropy, which is obtained by taking the orientation distribution entropy calculated based on the gray-level co-occurrence matrix; texture contrast, which is obtained by taking the contrast feature value of the gray-level co-occurrence matrix. Body proportion quantification: head-to-body ratio; Based on the quantized feature values ​​described above, construct the feature vector: ,in Let be the initial weights for each feature, and .

4. The system for identifying and warning of minors in special scenarios based on probability confidence as described in claim 3, characterized in that, The dynamic weight allocation of the calculation module follows the following rules: ； In the formula: The dynamic value of the i-th weight; Let be the inter-class variance of the i-th feature between the minor and non-minor samples. Similarly; Let be the within-class variance of the i-th feature in the sample of minors. Similarly; The preset time decay factor; This represents the time interval between the current moment and the last time the discriminative power of this feature was updated; After assigning dynamic weights, the feature vector V is compared with the dynamic weight vector. Perform a dot product to obtain a weighted eigenvector. Then The input is a probability model of minor features, which is used for adaptive matching. The output is the probability confidence value P.

5. The system for identifying and warning of minors in special scenarios based on probability confidence as described in claim 3, characterized in that, When the extraction module extracts body contour information, it performs edge detection on the collected body feature image data to obtain a continuous curve set of body contours, and then calculates the rate of curvature change of each curve segment in the curve set. Where K is the curvature of the curve segment and s is the arc length of the curve segment, in order to filter out the rate of change of curvature. Curve segments within a preset range are used as key body posture lines; At the same time, the length ratio of key body lines is calculated, that is, the ratio of the total length of key body lines to the total perimeter of the body contour, and then this ratio is added to the corresponding body line component in the feature vector V.

6. The system for identifying and warning of minors in special scenarios based on probability confidence as described in claim 1, characterized in that, The dynamic threshold adjustment process of the determination module is as follows: Obtain the current time period type M and real-time crowd density D of the current location, where M=1 corresponds to peak time, M=2 corresponds to off-peak time, and M=3 corresponds to low-peak time. Calculate the dynamic judgment threshold ; In the formula: The preset benchmark threshold is used for judgment. , The preset coefficient corresponding to time period type M; Real-time pedestrian density; Preset baseline pedestrian density; The expression for the probabilistic model of minor characteristics is: ; In the formula: The weighted cosine similarity between the feature vector and the reference vector is denoted as . , Represents the weighted eigenvector, This represents the baseline vector of the pre-trained minor feature model of the system. The L2 norm of a vector. Indicates the preset minimum value; As a dynamic weighted effectiveness factor, , represent the standard deviation of the dynamic weight vector Ω and the maximum value in the dynamic weight vector, respectively; The sample distribution adaptation factor is the ratio of the number of matches between the current feature vector and the same type of sample in the model to the total number of the same type of sample in the model. This is the distance attenuation coefficient; The probability confidence value P and the dynamic threshold Comparison: If P ≥ If P < 0, then it is determined to be a suspected minor; if P < 0. If so, it is determined that the individual is not a suspected minor.

7. The system for identifying and warning of minors in special scenarios based on probability confidence as described in claim 1, characterized in that, The warning level classification logic of the warning module is as follows: Obtain feature matching stability ,in Let N be the standard deviation of the probability confidence values ​​obtained from N consecutive data collections. The mean of the N confidence scores is the average of the N scores, where N represents the preset number of data collections. Calculate comprehensive early warning indicators ,in The preset adjustment coefficient, This represents the probability confidence level. when When the value is not less than the preset high-level threshold, it is a Level 1 warning; when... When the threshold is not less than the preset medium-level threshold and less than the preset high-level threshold, it is a level 2 warning; when When the value is less than the preset threshold, it is a Level 3 warning; The early warning module will output risk warning signals carrying P, S, Q and the warning level to the site management terminal.

8. The system for identifying and warning of minors in special scenarios based on probability confidence as described in claim 1, characterized in that, The sample library construction of the feedback module follows the following rules: The feature vector set verified as belonging to minors is divided into a facial feature subset and a body posture feature subset according to feature dimensions; For each subset, calculate the contribution of the feature dimension. ,in This represents the mean of that dimension in the dynamic weights. This represents the average contribution probability of this dimension in the matching operation; Based on contribution Build a hierarchical sample library for the index: use the feature dimensions with a contribution value higher than the preset contribution value threshold as the first-level index, and the rest as the second-level index.

9. The system for identifying and warning of minors in special scenarios based on probability confidence as described in claim 8, characterized in that, The feedback module's model feedback follows: Based on the index hierarchy of the sample library, feature samples corresponding to the first-level index are selected first for model parameter iteration. The iteration update formula is as follows: ; In the formula: These are the updated model parameters; These are the original parameters of the model; Set the learning rate; The gradient of the model's loss function; This represents the average contribution of the corresponding index-level feature dimension. in, The preset value range is [0.001, 0.1]. The value is larger when the parameters deviate greatly from the optimal solution in the early stage of model feedback and the average feature contribution is high. The value is smaller when the model approaches the optimal solution in the later stage, the average feature contribution is low, or the gradient of the loss function is small.

10. The system for identifying and warning of minors in special scenarios based on probability confidence as described in claim 1, characterized in that, The acquisition module is interconnected with the extraction module via a wireless network. The extraction module is interconnected with the computing module via a wireless network. The computing module is interconnected with the judgment module via a wireless network. The judgment module is interconnected with the early warning module via a wireless network. The early warning module is interconnected with the feedback module via a wireless network.

Images